From 93038ff7e993433948b3c09689c86b00aac971cf Mon Sep 17 00:00:00 2001 From: Simon Giraudot Date: Fri, 8 Jan 2016 09:03:56 +0100 Subject: [PATCH] More precisions about reconstruction methods principles --- .../Tutorials/Tutorial_reconstruction.txt | 27 +++++++++++++------ 1 file changed, 19 insertions(+), 8 deletions(-) diff --git a/Documentation/doc/Documentation/Tutorials/Tutorial_reconstruction.txt b/Documentation/doc/Documentation/Tutorials/Tutorial_reconstruction.txt index 9aa1c8eb97e..2ba15fb3d59 100644 --- a/Documentation/doc/Documentation/Tutorials/Tutorial_reconstruction.txt +++ b/Documentation/doc/Documentation/Tutorials/Tutorial_reconstruction.txt @@ -251,10 +251,14 @@ consistent. \subsection TutorialsReconstruction_reconstruction_poisson Poisson -Poisson reconstruction uses points with normals to produce smooth -closed surfaces. It is not indicated if the surface is expected to -pass exactly on the input points. On the contrary, it performs well if -the aim is to approximate a noisy point cloud with a smooth surface. +Poisson reconstruction consists in computing an indicator function +whose gradient matches the input normal vector field: this indicator +function has opposite signs inside and outside of the inferred shape +(hence the need for closed shapes). This method thus requires normals +and produces smooth closed surfaces. It is not appropriate if the +surface is expected to pass exactly on the input points. On the +contrary, it performs well if the aim is to approximate a noisy point +cloud with a smooth surface. Notice that it does not generate directly a mesh but computes an _implicit function_ (that can later be used to generate a mesh): @@ -308,7 +312,10 @@ be used: Advancing front is a Delaunay-based approach that generates triples of point indices that describe the triangular facets of the -reconstruction. Its main asset is to generate oriented manifold +reconstruction: it uses a priority queue to sequentially pick the +Delaunay facet the most likely to be part of the surface, based on a +size criterion (to favor the small facets) and an angle criterion (to +favor smoothness). Its main asset is to generate oriented manifold surfaces with boundaries: contrary to Poisson, it does not require normals and is not bound to reconstruct closed shapes. However, it requires preprocessing if the point is noisy. @@ -329,9 +336,13 @@ CGAL::advancing_front_surface_reconstruction(points.begin(), Scale space reconstruction aims at producing a surface that interpolates the input points (interpolant) while offering some -robustness to noise. It is the good choice if the input point cloud is -noisy but the user still wants the surface to pass exactly through the -points. +robustness to noise. More specifically, it first applies several times +a smoothing filter to the input point set to produce a scale space; +then, the smoothest scale is meshed using an alpha shape; finally, the +resulting connectivity between smoothed points is propagated to the +original raw input point set. This method is the right choice if the +input point cloud is noisy but the user still wants the surface to +pass exactly through the points. Notice that although there is an option to force the output to be manifold, it is not guaranteed to be orientable (contrary to _Poisson_